[GSD-13275] Regression: llama.cpp SYCL crashes with SIGSEGV on Intel Arc Pro B70 after upgrading compute runtime from 26.22.38646.6 to 26.27.39122.12

[settembre21]

Pre-submission Checklist

• I am using the latest GPU driver version (releases)
• I have searched for similar issues and found none

GPU Hardware

Intel ARC Pro B70

DRI Devices Information

$ ls -la /dev/dri/by-path/
total 0
drwxr-xr-x  2 root root 140 Aug 15 03:45 .
drwxr-xr-x  3 root root 140 Aug 15 03:45 ..
lrwxrwxrwx+ 1 root root   8 Aug 15 03:45 pci-0000:03:00.0-card -> ../card1
lrwxrwxrwx+ 1 root root  13 Aug 15 03:45 pci-0000:03:00.0-render -> ../renderD128
lrwxrwxrwx+ 1 root root   8 Aug 15 03:45 pci-0000:c9:00.0-card -> ../card2
lrwxrwxrwx  1 root root   8 Aug 15 03:45 pci-0000:c9:00.0-platform-simple-framebuffer.0-card -> ../card0
lrwxrwxrwx+ 1 root root  13 Aug 15 03:45 pci-0000:c9:00.0-render -> ../renderD129

GPU Detailed Information (lspci output)

$ sudo lspci -vvv -k -s 0000:03:00.0
[sudo: authenticate] Password:         
03:00.0 VGA compatible controller: Intel Corporation Battlemage G31 [Intel Graphics] (prog-if 00 [VGA controller])
	Subsystem: Sparkle Computer Co., Ltd. Device 0105
	Control: I/O- Mem+ BusMaster+ SpecCycle- MemWINV- VGASnoop- ParErr- Stepping- SERR- FastB2B- DisINTx-
	Status: Cap+ 66MHz- UDF- FastB2B- ParErr- DEVSEL=fast >TAbort- <TAbort- <MAbort- >SERR- <PERR- INTx-
	Latency: 0, Cache Line Size: 64 bytes
	Interrupt: pin ? routed to IRQ 135
	IOMMU group: 18
	Region 0: Memory at 2e00000000 (64-bit, prefetchable) [size=16M]
	Region 2: Memory at 1800000000 (64-bit, prefetchable) [size=32G]
	Expansion ROM at dc600000 [disabled] [size=2M]
	Capabilities: [40] Vendor Specific Information: Intel Capabilities v1
		CapA: Peg60Dis- Peg12Dis- Peg11Dis- Peg10Dis- PeLWUDis- DmiWidth=x4
		      EccDis- ForceEccEn- VTdDis- DmiG2Dis- PegG2Dis- DDRMaxSize=Unlimited
		      1NDis- CDDis- DDPCDis- X2APICEn- PDCDis- IGDis- CDID=0 CRID=0
		      DDROCCAP+ OCEn- DDRWrtVrefEn+ DDR3LEn+
		CapB: ImguDis- OCbySSKUCap- OCbySSKUEn- SMTCap- CacheSzCap 0x0
		      SoftBinCap- DDR3MaxFreqWithRef100=Disabled PegG3Dis-
		      PkgTyp- AddGfxEn- AddGfxCap- PegX16Dis- DmiG3Dis- GmmDis-
		      DDR3MaxFreq=2932MHz LPDDR3En-
	Capabilities: [70] Express (v2) Endpoint, IntMsgNum 0
		DevCap:	MaxPayload 256 bytes, PhantFunc 0, Latency L0s unlimited, L1 unlimited
			ExtTag+ AttnBtn- AttnInd- PwrInd- RBE+ FLReset+ SlotPowerLimit 0W TEE-IO-
		DevCtl:	CorrErr- NonFatalErr- FatalErr- UnsupReq-
			RlxdOrd+ ExtTag+ PhantFunc- AuxPwr- NoSnoop+ FLReset-
			MaxPayload 256 bytes, MaxReadReq 512 bytes
		DevSta:	CorrErr- NonFatalErr- FatalErr- UnsupReq- AuxPwr- TransPend-
		LnkCap:	Port #0, Speed 2.5GT/s, Width x1, ASPM L0s L1, Exit Latency L0s <64ns, L1 <1us
			ClockPM- Surprise- LLActRep- BwNot- ASPMOptComp+
		LnkCtl:	ASPM Disabled; RCB 64 bytes, LnkDisable- CommClk-
			ExtSynch- ClockPM- AutWidDis- BWInt- AutBWInt- FltModeDis-
		LnkSta:	Speed 2.5GT/s, Width x1
			TrErr- Train- SlotClk- DLActive- BWMgmt- ABWMgmt-
		DevCap2: Completion Timeout: Range B, TimeoutDis+ NROPrPrP- LTR+
			 10BitTagComp+ 10BitTagReq+ OBFF Not Supported, ExtFmt+ EETLPPrefix-
			 EmergencyPowerReduction Not Supported, EmergencyPowerReductionInit-
			 FRS- TPHComp- ExtTPHComp-
			 AtomicOpsCap: 32bit- 64bit- 128bitCAS-
		DevCtl2: Completion Timeout: 50us to 50ms, TimeoutDis-
			 AtomicOpsCtl: ReqEn-
			 IDOReq- IDOCompl- LTR+ EmergencyPowerReductionReq-
			 10BitTagReq- OBFF Disabled, EETLPPrefixBlk-
		LnkCap2: Supported Link Speeds: 2.5GT/s, Crosslink- Retimer- 2Retimers- DRS-
		LnkCtl2: Target Link Speed: 2.5GT/s, EnterCompliance- SpeedDis-
			 Transmit Margin: Normal Operating Range, EnterModifiedCompliance- ComplianceSOS-
			 Compliance Preset/De-emphasis: -6dB de-emphasis, 0dB preshoot
		LnkSta2: Current De-emphasis Level: -6dB, EqualizationComplete- EqualizationPhase1-
			 EqualizationPhase2- EqualizationPhase3- LinkEqualizationRequest-
			 Retimer- 2Retimers- CrosslinkRes: unsupported, FltMode-
	Capabilities: [ac] MSI: Enable+ Count=1/1 Maskable+ 64bit+
		Address: 00000000fee00000  Data: 0000
		Masking: 00000000  Pending: 00000000
	Capabilities: [d0] Power Management version 3
		Flags: PMEClk- DSI- D1- D2- AuxCurrent=0mA PME(D0+,D1-,D2-,D3hot+,D3cold-)
		Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME-
	Capabilities: [100 v1] Alternative Routing-ID Interpretation (ARI)
		ARICap:	MFVC- ACS-, Next Function: 0
		ARICtl:	MFVC- ACS-, Function Group: 0
	Capabilities: [110 v1] Null
	Capabilities: [200 v1] Address Translation Service (ATS)
		ATSCap:	Invalidate Queue Depth: 00
		ATSCtl:	Enable+, Smallest Translation Unit: 00
	Capabilities: [420 v1] Physical Resizable BAR
		BAR 2: current size: 32GB, supported: 256MB 512MB 1GB 2GB 4GB 8GB 16GB 32GB
	Capabilities: [220 v1] Virtual Resizable BAR
		BAR 2: current size: 8GB, supported: 256MB 512MB 1GB 2GB 4GB 8GB 16GB 32GB
	Capabilities: [320 v1] Single Root I/O Virtualization (SR-IOV)
		IOVCap:	Migration- 10BitTagReq+ IntMsgNum 0
		IOVCtl:	Enable- Migration- Interrupt- MSE- ARIHierarchy+ 10BitTagReq-
		IOVSta:	Migration-
		Initial VFs: 7, Total VFs: 7, Number of VFs: 0, Function Dependency Link: 00
		VF offset: 1, stride: 1, Device ID: e223
		Supported Page Size: 00000553, System Page Size: 00000001
		Region 0: Memory at 0000002e01000000 (64-bit, prefetchable)
		Region 2: Memory at 0000002000000000 (64-bit, prefetchable)
		VF Migration: offset: 00000000, BIR: 0
	Capabilities: [400 v1] Latency Tolerance Reporting
		Max snoop latency: 1048576ns
		Max no snoop latency: 1048576ns
	Kernel driver in use: xe
	Kernel modules: xe

Driver Version

26.27.39122.12-1~26.04~ppa1
(Kobuk Team PPA package; based on upstream 26.27.39122.11)

Installed GPU Driver Packages

$ sudo dpkg --list | grep -iE " igc|gmm|opencl|level-zero|fc|level_zero|ocloc|libze "
ii  bpfcc-tools                                           0.35.0+ds-1ubuntu2                          all          tools for BPF Compiler Collection (BCC)
ii  clinfo                                                3.0.25.02.14-1build1                        amd64        Query OpenCL system information
ii  fcitx-mozc-data                                       2.29.5160.102+dfsg-1.4ubuntu5               all          Mozc input method - data files for fcitx
ii  fcitx5                                                5.1.19-1                                    amd64        Fcitx Input Method Framework v5
ii  fcitx5-config-qt                                      5.1.13-1                                    amd64        configuration tool for Fcitx5 (Qt version)
ii  fcitx5-data                                           5.1.19-1                                    all          Fcitx Input Method Framework v5 (common data files)
ii  fcitx5-frontend-all                                   5.1.19-1                                    all          Fcitx Input Method Framework v5 (Frontend IM Module metapackage)
ii  fcitx5-frontend-gtk3                                  5.1.6-1                                     amd64        GTK3 IM Module for fcitx5
ii  fcitx5-frontend-gtk4                                  5.1.6-1                                     amd64        GTK4 IM Module for fcitx5
ii  fcitx5-frontend-qt5                                   5.1.13-1                                    amd64        Qt5 IM module for fcitx5
ii  fcitx5-frontend-qt6                                   5.1.13-1                                    amd64        Qt6 IM module for fcitx5
ii  fcitx5-modules:amd64                                  5.1.19-1                                    amd64        Fcitx Input Method Framework v5 (core modules)
ii  fcitx5-mozc:amd64                                     2.29.5160.102+dfsg-1.4ubuntu5               amd64        Mozc engine for fcitx5 - Client of the Mozc input method
hi  intel-ocloc                                           26.22.38646.6-1~25.10~ppa1                  amd64        Tool for managing Intel Compute GPU device binary format
hi  intel-opencl-icd                                      26.22.38646.6-1~25.10~ppa1                  amd64        Intel graphics compute runtime for OpenCL
ii  libbpfcc:amd64                                        0.35.0+ds-1ubuntu2                          amd64        shared library for BPF Compiler Collection (BCC)
ii  libcbor0.10:amd64                                     0.10.2-2ubuntu3                             amd64        library for parsing and generating CBOR (RFC 7049)
ii  libdebconfclient0:amd64                               0.280ubuntu1                                amd64        Debian Configuration Management System (C-implementation library)
ii  libfcitx5-qt-data                                     5.1.13-1                                    all          Qt library and IM module for fcitx5 (data files)
ii  libfcitx5-qt1:amd64                                   5.1.13-1                                    amd64        Qt library and IM module for fcitx5
ii  libfcitx5-qt6-1:amd64                                 5.1.13-1                                    amd64        Qt6 library and IM module for fcitx5
ii  libfcitx5config6:amd64                                5.1.19-1                                    amd64        Fcitx Input Method Framework v5 (config library)
ii  libfcitx5core7:amd64                                  5.1.19-1                                    amd64        Fcitx Input Method Framework v5 (core library)
ii  libfcitx5gclient2:amd64                               5.1.6-1                                     amd64        GLib-based D-Bus client library for fcitx5 (library)
ii  libfcitx5utils2:amd64                                 5.1.19-1                                    amd64        Fcitx Input Method Framework v5 (utils library)
ii  libfile-fcntllock-perl                                0.22-4ubuntu6                               amd64        Perl module for file locking with fcntl(2)
ii  libigc2                                               2.38.3-3~26.04                              amd64        Core libraries for Intel(R) Graphics Compiler for OpenCL(TM)
ii  libigdfcl2                                            2.38.3-3~26.04                              amd64        OpenCL library for Intel(R) Graphics Compiler for OpenCL(TM)
ii  libigdgmm12:amd64                                     22.10.0-1~26.04~ppa1                        amd64        Intel Graphics Memory Management Library -- shared library
ii  libref-array1t64:amd64                                0.6.2-3build1                               amd64        refcounted array for C
ii  linux-firmware-qlogic                                 20260319.git217ca6e4-0ubuntu1.1             all          Firmware for QLogic SCSI, FC, and IB host bus and Ethernet adapters
ii  node-punycode                                         2.3.1+~2.1.4-1                              all          Nodejs robust Punycode converter fully RFC compliant
ii  node-uuid                                             8.3.2+~8.3.4-3                              all          simple and fast RFC4122 UUID generation - Node.js module
ii  node-ws                                               8.19.0+~cs14.19.1-1                         all          RFC-6455 WebSocket implementation module for Node.js
ii  ocl-icd-libopencl1:amd64                              2.3.4-1                                     amd64        Generic OpenCL ICD Loader
ii  python3-bpfcc                                         0.35.0+ds-1ubuntu2                          all          Python 3 wrappers for BPF Compiler Collection (BCC)
ii  python3-idna                                          3.11-1ubuntu0.1                             all          Python IDNA2008 (RFC 5891) handling (Python 3)

Note: The installed-package list above was captured after rolling back to the
last known working version. The crash was reproduced while libze-intel-gpu1,
intel-opencl-icd, and intel-ocloc were installed at
26.27.39122.12-1~26.04~ppa1.

Driver Installation Details

sudo add-apt-repository ppa:kobuk-team/intel-graphics
sudo apt update
sudo apt install libze-intel-gpu1 intel-opencl-icd intel-ocloc

Linux Distribution

Other (please specify below)

Other Linux Distribution

Ubuntu 26.04 LTS

Kernel Version & Boot Parameters

$ uname -r
7.0.0-29-generic
$ cat /proc/cmdline
BOOT_IMAGE=/boot/vmlinuz-7.0.0-29-generic root=UUID=9177ea2b-d319-4782-819c-fb89019bd458 ro quiet splash crashkernel=2G-4G:320M,4G-32G:512M,32G-64G:1024M,64G-128G:2048M,128G-:4096M
$ lsmod | grep -E ' i915|xe '
xe                   4419584  9

Actual Behavior

This was reproduced with the Ubuntu 26.04 packages from the Kobuk Team Intel Graphics PPA. The packaged version is 26.27.39122.12-1~26.04~ppa1. I have not yet reproduced it using Intel's official 26.27.39122.11 release packages.

Description

I found a reproducible regression in the Intel GPU compute runtime packages
provided by the Kobuk Team Intel Graphics PPA for Ubuntu 26.04.

llama.cpp SYCL workloads run successfully with:

• libze-intel-gpu1: 26.22.38646.6-1~25.10~ppa1
• intel-opencl-icd: 26.22.38646.6-1~25.10~ppa1
• intel-ocloc: 26.22.38646.6-1~25.10~ppa1

After upgrading all three packages to:

• libze-intel-gpu1: 26.27.39122.12-1~26.04~ppa1
• intel-opencl-icd: 26.27.39122.12-1~26.04~ppa1
• intel-ocloc: 26.27.39122.12-1~26.04~ppa1

the first llama.cpp SYCL benchmark crashes with SIGSEGV.

The regression was reproduced twice. Downgrading the three packages restores
normal operation.

System information

• OS: Ubuntu 26.04 LTS
• Kernel: 7.0.0-29-generic
• Architecture: x86_64
• CPU: AMD Ryzen 7 8845HS
• GPU: Intel Arc Pro B70 / Battlemage G31
• PCI ID: 8086:e223
• GPU subsystem: 172f:0105
• Kernel driver: xe
• System memory: 53 GiB
• oneAPI DPC++ compiler:
  Intel oneAPI DPC++/C++ Compiler 2026.1.1
  Package version: 2026.1.1-325
• llama.cpp build: b10438
• llama.cpp commit:
  9d57ce456c94d241dde672b2db9cf18879766568

The runtime packages were installed from:

https://ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu

Build configuration

llama.cpp was built with the following relevant options:

CMAKE_BUILD_TYPE=Release
GGML_NATIVE=ON
GGML_SYCL=ON
GGML_SYCL_F16=ON
GGML_SYCL_DNN=ON
GGML_SYCL_GRAPH=ON
GGML_SYCL_HOST_MEM_FALLBACK=ON
GGML_SYCL_SUPPORT_LEVEL_ZERO_API=ON

The compiler was:

Intel(R) oneAPI DPC++/C++ Compiler 2026.1.1

Runtime environment

export ONEAPI_DEVICE_SELECTOR=level_zero:gpu
export GGML_SYCL_ENABLE_OPT=1
export GGML_SYCL_USE_LEVEL_ZERO_API=1
export GGML_SYCL_ENABLE_VMM=1
export GGML_SYCL_ENABLE_DNN=1
export GGML_SYCL_ENABLE_FLASH_ATTN=1
export GGML_SYCL_ENABLE_GRAPH=0
export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1

Reproduction command

llama-bench
-m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf
-ngl 999
-dev SYCL0
-t 8
-ctk q8_0
-ctv q8_0
-fa on
-sm none
-p 512
-n 0
-r 1
-o json

Actual result

The process crashes before producing benchmark results:

Segmentation fault (core dumped)
Exit status: 139

Kernel log:

llama-bench: segfault at 0 ip 0000000000000000
error 14 likely on CPU 1 (core 1, socket 0)

The same failure occurred in two independent test runs.

Expected result

The benchmark should complete normally, as it does with
26.22.38646.6-1~25.10~ppa1.

With the older package set and the same llama.cpp commit, compiler, model and
benchmark parameters, I obtained:

26B prompt processing, 512 tokens: 1396.45 tokens/s
26B text generation, 128 tokens: 56.75 tokens/s
31B prompt processing, 512 tokens: 462.83 tokens/s
31B text generation, 128 tokens: 21.99 tokens/s

The API test using the same SYCL backend also completed successfully.

Regression verification

Test procedure:

1. Install 26.22.38646.6 packages.
2. Build llama.cpp b10438.
3. Run the benchmark suite: PASS.
4. Upgrade all three GPU runtime packages to 26.27.39122.12.
5. Rebuild the same llama.cpp commit with the same oneAPI compiler.
6. Run the first 26B PP512 benchmark: SIGSEGV, exit 139.
7. Downgrade all three packages to 26.22.38646.6.
8. Load the production 31B model and check the health endpoint: PASS.

The package upgrade/downgrade and failure were reproduced twice.

Workaround

Downgrade and hold the following packages at
26.22.38646.6-1~25.10~ppa1:

libze-intel-gpu1
intel-opencl-icd
intel-ocloc

Additional notes

All three packages were upgraded together, so I have not yet isolated whether
the regression is specifically in libze-intel-gpu1, intel-opencl-icd, ocloc
generated device code, or an interaction between these components.

The workload explicitly selects the Level Zero backend.

Ubuntu also displayed an Apport notification saying that timeout crashed
with SIGSEGV. This was a secondary effect: the Rust implementation of
timeout mirrored the child process signal. The kernel log confirms that the
original crashing process was llama-bench.

Please let me know if you need:

• the full llama.cpp build log;
• llama-server --list-devices output;
• the exact package manifests;
• the benchmark logs;
• a stack trace or debug-symbol reproduction;
• testing with the official upstream 26.27.39122.11 packages.

Expected Behavior

The llama.cpp SYCL benchmark should complete successfully without crashing,
as it does with the previous compute runtime version
26.22.38646.6-1~25.10~ppa1.

Upgrading to 26.27.39122.12-1~26.04~ppa1 should preserve compatibility with
Intel Arc Pro B70 and produce valid benchmark results when using the Level Zero
backend.

Reproduction Rate

Always reproduces - 100%

Steps to Reproduce

Steps to reproduce

1. On Ubuntu 26.04 with an Intel Arc Pro B70, enable the Kobuk Team Intel
   Graphics PPA:

   sudo add-apt-repository ppa:kobuk-team/intel-graphics
   sudo apt update

2. Install the affected Intel GPU compute runtime packages:

   sudo apt install
   libze-intel-gpu1=26.27.39122.12-1~26.04~ppa1
   intel-opencl-icd=26.27.39122.12-1~26.04~ppa1
   intel-ocloc=26.27.39122.12-1~26.04~ppa1

3. Build llama.cpp commit
   9d57ce456c94d241dde672b2db9cf18879766568 with Intel oneAPI DPC++/C++
   Compiler 2026.1.1 and the following CMake options:

   cmake -S . -B build -G Ninja
   -DCMAKE_BUILD_TYPE=Release
   -DCMAKE_C_COMPILER=/opt/intel/oneapi/2026.1/bin/icx
   -DCMAKE_CXX_COMPILER=/opt/intel/oneapi/2026.1/bin/icpx
   -DBUILD_SHARED_LIBS=ON
   -DGGML_NATIVE=ON
   -DGGML_SYCL=ON
   -DGGML_SYCL_F16=ON
   -DGGML_SYCL_DNN=ON
   -DGGML_SYCL_GRAPH=ON
   -DGGML_SYCL_HOST_MEM_FALLBACK=ON
   -DGGML_SYCL_SUPPORT_LEVEL_ZERO_API=ON

   cmake --build build --parallel

4. Set the SYCL runtime environment:

   source /opt/intel/oneapi/2026.1/oneapi-vars.sh --force

   export ONEAPI_DEVICE_SELECTOR=level_zero:gpu
   export GGML_SYCL_ENABLE_OPT=1
   export GGML_SYCL_USE_LEVEL_ZERO_API=1
   export GGML_SYCL_ENABLE_VMM=1
   export GGML_SYCL_ENABLE_DNN=1
   export GGML_SYCL_ENABLE_FLASH_ATTN=1
   export GGML_SYCL_ENABLE_GRAPH=0
   export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1

5. Run the benchmark using a Gemma 4 26B GGUF model:

   ./build/bin/llama-bench
   -m /path/to/gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf
   -ngl 999
   -dev SYCL0
   -t 8
   -ctk q8_0
   -ctv q8_0
   -fa on
   -sm none
   -p 512
   -n 0
   -r 1
   -o json

6. Observe that llama-bench terminates with SIGSEGV before producing valid
   benchmark output:

   Segmentation fault (core dumped)
   Exit status: 139

7. Downgrade all three runtime packages to
   26.22.38646.6-1~25.10~ppa1, rebuild the same llama.cpp commit, and repeat
   the benchmark. It completes successfully under otherwise identical
   conditions.

Is this a regression?

• Yes, this is a regression - functionality that previously worked is now broken

Last Known Working Driver Version

26.22.38646.7-1~26.04~ppa1

First Known Failing Driver Version

26.27.39122.12-1~26.04~ppa1

API Call Logs

No response

strace Logs

No response

System Logs / dmesg Output

No response

Backtrace (if crash or hang occurred)

No response

Source Code / Reproducer

No response

Command Line / Application Details

No response

oneAPI Version (if applicable)

$ icpx --version
Intel(R) oneAPI DPC++/C++ Compiler 2026.1.1 (2026.1.1.20260724)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /opt/intel/oneapi/compiler/2026.1/bin/compiler
Configuration file: /opt/intel/oneapi/compiler/2026.1/bin/compiler/../icpx.cfg

Screenshots / Video

No response

Additional Notes

No response

[kgibala]

Hi @settembre21,

Thank you for your contribution and for reporting this regression.

We will investigate it internally and provide an update as soon as we have something meaningful to share.

Thanks.